def get_data_for_auto_init(is_upper, bbox_list, test_img_idx):
img = cv2.imread('Data/Radiographs/'+str(test_img_idx).zfill(2)+'.tif')
[mean_bbox, search_region] = get_parameters(is_upper, bbox_list, img)
def_width = abs(mean_bbox[0] - mean_bbox[2])
def_height = abs(mean_bbox[1] - mean_bbox[3])
smallImages = np.zeros((13, def_width * def_height)) # building model excluding test image
# # can load preprocessed images directly
# radiographs = task2.load(preprocessed=True)
# or can load and then preprocess
radiographs = task2.load(preprocessed=False)
del radiographs[test_img_idx-1] # deleteing test index
# skip_amf = without median filter
radiographs = [task2.enhance(radiograph, skip_amf=True) for radiograph in radiographs]
for ind, radiograph in enumerate(radiographs):
[x1, y1, x2, y2] = bbox_list[ind]
cutImage = radiograph[y1:y2, x1:x2]
result = cv2.resize(cutImage, (def_width, def_height), interpolation=cv2.INTER_NEAREST)
smallImages[ind] = result.flatten()
return [smallImages, [def_width, def_height, search_region]]
def buildASM(incisor_list, test_img_idx, k):
"""
Build ASM of a given incisor list.
Excludes the test image from model construction.
Args:
k: No. of pixels on either side of a model point for grey level model
"""
asm_list = []
train_imgs = task2.load()
del train_imgs[test_img_idx-1]
with Timer("Building Active Shape Model"):
for incisor in incisor_list:
print("..For Incisor "+str(incisor))
train_lms = load_all_landmarks_of(incisor, test_img_idx)
directory = "ASM_Models/test_img_%02d/" %(test_img_idx)
filename = "incisor_%d.model" %(incisor)
if not os.path.exists(directory+filename):
asm = ASM(incisor, train_lms, train_imgs, k)
asm_list.append(asm)
save_asm(asm, directory, filename)
else:
with file(directory+filename, 'rb') as f:
asm_list.append( pickle.load(f) )
print("") # just for elegant printing on screen
return asm_list
def fit_model(asm_list,
incisor_list,
test_img_idx,
m,
auto_estimate=True,
save=False,
show=False):
global save_plots
global test_idx
global show_plots
save_plots = save
test_idx = test_img_idx
show_plots = show
test_img = task2.load([test_img_idx])[0]
X_init_list = []
if auto_estimate:
X_init_list = auto_init.get_estimate(asm_list, incisor_list,
test_img_idx)
else:
lms_list = []
for asm in asm_list:
lms_list.append(asm.sm.mean_shape)
X_init_list = manual_init.get_estimate(lms_list, incisor_list,
test_img)
Plots.plot_landmarks_on_image(X_init_list, test_img, title="manual_init",\
show=False, save=False, wait=True, color=(0,255,0))
with Timer("Fitting Model in Multi Resolution Framework"):
final_fit = []
for ind, X in enumerate(X_init_list):
print("..For incisor %d" % (asm_list[ind].incisor))
pyramid = task1.gauss_pyramid(test_img, PYRAMID_LEVELS)
X = X.scaleposition(1.0 / 2**(PYRAMID_LEVELS + 1))
level = PYRAMID_LEVELS
for img in reversed(pyramid):
#print("..Level %d" %(level))
X = X.scaleposition(2)
X = fit_one_level(X, img, asm_list[ind], level, m, MAX_ITER)
level -= 1
final_fit.append(X)
print("") # just for elegant printing on screen
if save_plots or show_plots:
directory = "Plots/model_fit/test_img_" + str(test_idx) + "/"
Plots.plot_landmarks_on_image(final_fit, test_img, directory=directory, title="Final_fit",\
show=show_plots, save=save_plots, wait=True, color=(0,255,0))
return final_fit
def test_auto_initial_estimate(incisor_list, test_img_idx, k):
asm_list = task1.buildASM(incisor_list, test_img_idx, k)
X_init_list = auto_init.get_estimate(asm_list, incisor_list, test_img_idx, show_bbox_dist=True, \
show_app_models=False, show_finding_bbox=True, \
show_autoinit_bbox=False, show_autoinit_lms=False, save=False)
test_img = task2.load([test_img_idx])[0]
directory = "Plots/auto_init/test_img_%02d/" % (test_img_idx)
Plots.plot_landmarks_on_image(X_init_list, test_img, directory=directory, title="auto_init",\
show=False, save=True, wait=True, color=(0,255,0))
def test_manual_init(incisor_list, test_img_idx):
asm_list = task1.buildASM(incisor_list, test_img_idx, k)
lms_list = []
for asm in asm_list:
lms_list.append(asm.sm.mean_shape)
test_img = task2.load([test_img_idx])[0]
X_init_list = manual_init.get_estimate(lms_list, test_img)
Plots.plot_landmarks_on_image(X_init_list, test_img, title="manual_init",\
show=True, save=False, wait=True, color=(0,255,0))
def get_big_bbox(is_upper, test_img_idx):
"""
Finds the bounding box surrounding all the four upper(or lower) incisors
"""
bbox_list = extract_roi_for_appModel(is_upper, test_img_idx)
directory = "Autoinit_params/test_img_%02d/" %(test_img_idx)
filename = "upper_incisor.model" if is_upper else "lower_incisor.model"
if os.path.exists(directory+filename):
with file(directory+filename, 'rb') as f:
[smallImages, [def_width, def_height, search_region]] = pickle.load(f)
else:
[smallImages, [def_width, def_height, search_region]] = \
get_data_for_auto_init(is_upper, bbox_list, test_img_idx)
save_file([smallImages, [def_width, def_height, search_region]], directory, filename)
[_, evecs, mean] = pca(smallImages, 5)
global jaw_split
global plot_app_models
global save_plots
global save_dir
#Visualize the appearance model
app_model = np.hstack( (mean.reshape(def_height,def_width), \
normalize(evecs[:,0].reshape(def_height,def_width)), \
normalize(evecs[:,1].reshape(def_height,def_width)), \
normalize(evecs[:,2].reshape(def_height,def_width))) \
).astype(np.uint8)
if plot_app_models:
cv2.imshow('app_model', app_model)
cv2.waitKey(0)
if save_plots:
title = "upper_incisors" if is_upper else "lower_incisors"
save_image(app_model, "appearance_model_"+title+".png", save_dir)
test_img = task2.load([test_img_idx])[0]
test_img = task2.enhance(test_img, skip_amf=True)
[(a, b), (c, d)] = find_bbox(mean, evecs, test_img, def_width, def_height, is_upper, \
jaw_split, search_region)
return [(a, b), (c, d)]
def evaluate_results(test_img_idx, incisor_list, final_fit_list):
"""
Uses the Dice Coefficient to evaulate the similarity between the final landmarks
given by the model and the ground truth landmarks of the test image
"""
test_img = task2.load([test_img_idx])[0]
test_lms_list = load_landmarks(test_img_idx, incisor_list)
dice_scores = []
height, width, _ = test_img.shape
with Timer("Evaluating Results"):
print("..Dice similarity score")
for ind, incisor in enumerate(incisor_list):
image1 = np.zeros((height, width), np.uint8)
image2 = np.zeros((height, width), np.uint8)
X1 = test_lms_list[ind].as_matrix() # X1 - ground truth
cv2.fillPoly(image1, np.int32([X1]), 255)
X2 = final_fit_list[ind].as_matrix() # X2 - best fit
cv2.fillPoly(image2, np.int32([X2]), 255)
dice = np.sum(image1[image2 == 255]) * 2.0 / (np.sum(image1) +
np.sum(image2))
print("....for incisor %02d - %.2f" % (incisor, dice))
dice_scores.append(dice)
if save_plots or show_plots:
Plots.plot_results(np.arange(len(dice_scores)), incisor_list, dice_scores, \
test_img_idx, show=show_plots, save=save_plots)
def get_estimate(asm_list,incisor_list, test_img_idx, show_bbox_dist=False, show_app_models=False, \
show_finding_bbox=False, show_autoinit_bbox=False, show_autoinit_lms=False, save=False):
"""
Finds an initial estimate for all the incisors in the incisor_list
"""
global plot_bbox_dist
global plot_app_models
global plot_finding_bbox
global plot_autoinit_bbox
global plot_autoinit_lms
global save_plots
global save_dir
global jaw_split
plot_bbox_dist = show_bbox_dist
plot_app_models = show_app_models
plot_finding_bbox = show_finding_bbox
plot_autoinit_bbox = show_autoinit_bbox
plot_autoinit_lms = show_autoinit_lms
save_plots = save
save_dir = "Plots/auto_init/test_img_%02d/" %(test_img_idx)
with Timer("Finding Initial Estimate automatically"):
if any(incisor < 5 for incisor in incisor_list): # upper incisor
is_upper= True
with Timer("..for upper incisors", dots="...."):
[(w1U, h1U), (w2U, h2U)] = get_big_bbox(is_upper, test_img_idx)
if any(incisor > 4 for incisor in incisor_list): # lower incisor
is_upper= False
with Timer("..for lower incisors", dots="...."):
[(w1L, h1L), (w2L, h2L)] = get_big_bbox(is_upper, test_img_idx)
print("") # just for elegant printing on screen
init_list = []
test_img = task2.load([test_img_idx])[0]
img_org = test_img.copy()
test_img = task2.enhance(test_img, skip_amf=True)
for index,incisor in enumerate(incisor_list):
# Assume all teeth have more or less the same width
if incisor < 5:
ind = incisor
bbox = [(w1U +(ind-1)*(w2U-w1U)/4, h1U), (w1U +(ind)*(w2U-w1U)/4, h2U)]
else:
ind = incisor - 4
bbox = [(w1L +(ind-1)*(w2L-w1L)/4, h1L), (w1L +(ind)*(w2L-w1L)/4, h2L)]
center = np.mean(bbox, axis=0)
Plots.plot_autoinit(test_img, jaw_split, lowest_error_bbox=bbox, directory=save_dir, \
title="initial_estimate_bbox_incisor_%d" %(incisor), wait=True, \
show=plot_autoinit_bbox, save=False)#save=save_plots
init = asm_list[index].sm.mean_shape.scale_to_bbox(bbox).translate(center)
Plots.plot_landmarks_on_image([init], img_org, directory=save_dir, \
title="initial_estimate_lms_incisor_%d" %(incisor), \
show=plot_autoinit_lms, save=False, color=(0,255,0))#save=save_plots
init_list.append(init)
return init_list